Full text PDF

Regularized regressions are widely used in psychological research where the fitted model is assumed to be correctly specified. Yet, psychological data routinely violates assumptions of linearity, homoscedasticity and additivity which raises questions about estimator robustness when models are misspecified. The present simulation study examined how regularized estimators—ridge, LASSO, elastic net, adaptive LASSO, SCAD and MCP—degrade under compound misspecification. Using high-dimensional data with block-wise correlation (p = 2,000) and three sample sizes (n = 500, 2,000 and 10,000), we generated outcomes under correct specification and compound misspecification incorporating nonlinearity, heteroscedasticity and omitted interactions. Performance was evaluated through predictive accuracy (out-of-sample R² and RMSE) and selection stability (mean Jaccard index across replicates). Results revealed that misspecification produced modest degradation in predictive accuracy with R-square reductions of 0.006–0.017 across sample sizes. Selection stability was minimally affected by misspecification at smaller samples. At n = 10,000, non-convex methods achieved highest stability.

Pages © TQMP;
Website last modified: 2026-09-20.
Template last modified: 2022-03-04 18h27.
Page consulted on .
Be informed of the upcoming issues with RSS feed: RSS